Leading AI Researchers Flag Challenges in Real-World Agent Deployment

A packed audience of students, researchers, and industry professionals gathered at the University of California, Berkeley, for the Agentic AI Summit, highlighting the intense interest in AI agents. The event brought together leading figures in the field, including Jakob Pachocki of OpenAI, Ed Chi of Google DeepMind, Bill Dally of Nvidia, and Ion Stoica of Databricks. The summit focused on AI agents—systems designed to perform tasks autonomously using various tools—yet the overall message was tempered with caution.

Despite the excitement, speakers acknowledged that the technology is still far from its full potential. Ed Chi from Google DeepMind emphasized the limitations of current AI agents, noting a significant gap between performance in controlled demonstrations and real-world applications. Pachocki echoed these concerns, raising red flags about the safety, security, and trustworthiness of agentic systems, especially as they begin to integrate into critical sectors.

Sherwin Wu, head of engineering at OpenAI API, admitted that the daily impact of AI agents on his work has been minimal, stating, “I still don’t think agents have really lived up to their promise.” This sentiment was shared by many attendees, who pointed to reliability issues such as agents failing to retain context or handle complex, multi-step tasks consistently.

However, the event was not without optimism. Ion Stoica highlighted improvements in infrastructure that are facilitating the development of more robust agentic systems. Bill Dally from Nvidia suggested that continued advancements in hardware would enable more sophisticated and efficient agent behavior. Additionally, several presenters noted “narrow wins” in specialized areas like coding, indicating progress even in the face of broader challenges.

The Agentic AI Summit underscored the industry’s long-term vision: to build AI agents capable of operating reliably in the real world. While the path is fraught with challenges, the potential rewards—ranging from increased productivity to transformative automation—make the pursuit worthwhile. For now, the focus remains on refining current systems and addressing the limitations that prevent agents from fulfilling their full promise.

As the field moves forward, the collaboration between leading research institutions and tech firms will be critical. OpenAI’s Sam Altman previously suggested that AI agents could begin “joining the workforce” in 2025, but the cautious outlook from top researchers indicates that such a transition will require significant advancements in both technology and infrastructure.

Source: [1]title:From OpenAI to Nvidia, researchers agree: AI agents have a long way to go

(url:https://fortune.com/2025/08/05/from-openai-to-nvidia-researchers-agree-ai-agents-have-a-long-way-to-go/)